PyAPX: Python toolkit for atomic configuration pattern exploration

This paper introduces PyAPX, a Python toolkit designed to accelerate materials discovery by performing Bayesian searches for stable atomic configurations using novel encoding methods that outperform traditional one-hot encoding, as demonstrated in the h-BCN system.

The Gist

Imagine you are a master chef trying to create the perfect dish. You have a recipe that tells you exactly which ingredients to use (the chemical composition) and the type of pot you need (the crystal structure). But here's the catch: even with the exact same ingredients and pot, the taste of the dish changes drastically depending on how you arrange the ingredients on the plate.

If you pile all the salt in one corner, it's too salty. If you scatter the herbs evenly, it's perfect. In the world of materials science, these "ingredients" are atoms, and their arrangement is called an atomic configuration.

This paper introduces a new digital tool called PyAPX (Python Atomic Pattern eXplorer) designed to help scientists find the "perfect plating" for new materials.

Here is a simple breakdown of what the researchers did and why it matters:

1. The Problem: The "Arrangement" Mystery

Scientists have been great at two things:

• Finding new pots: Predicting what stable crystal structures exist for a given mix of elements.
• Swapping ingredients: Figuring out which elements to swap in to get better properties (like making a battery last longer).

But they often struggle with the third step: Arranging the ingredients. Even if you know you need 6 Carbon, 6 Boron, and 6 Nitrogen atoms, there are millions of ways to arrange them. Some arrangements make the material a great insulator; others make it a superconductor. Finding the best arrangement by trying every possibility is like trying to find a specific needle in a haystack by checking every single straw one by one—it takes too long.

2. The Solution: PyAPX (The Smart Taster)

The authors built PyAPX, a software toolkit that acts like a super-smart, fast-learning taster. Instead of tasting every single possible arrangement (which would take forever), it uses a technique called Bayesian Optimization.

Think of it like this:

• The Old Way: You taste 100 random dishes, write down the scores, and guess which one to try next.
• The PyAPX Way: It tastes a few dishes, learns the "flavor profile," and then uses math to guess, "Based on what I just tasted, the dish with the salt in the middle-left is probably the best. Let's try that one next." It balances exploring new, weird combinations with exploiting the patterns it has already found.

3. The Secret Sauce: Better "Language" for Atoms

The biggest innovation in this paper isn't just the tool, but how it "speaks" to the computer. To make the AI learn, you have to describe the arrangement of atoms using numbers (encoding).

• The Old Language (One-Hot Encoding): Imagine describing a chessboard by just saying, "Square A1 has a Pawn." It tells you what is there, but not what is around it. It's like describing a person by only saying their name, ignoring their friends and family.
• The New Language (NA and NAmod Encoding): The authors created a new way to describe atoms that includes their neighbors. It's like saying, "Square A1 has a Pawn, and it is surrounded by three Knights."
  • They even added a "Modified" version that accounts for anisotropy (direction). This is like realizing that having a Knight to your left feels different than having a Knight to your right. This extra detail helps the AI understand the "vibe" of the atomic neighborhood much better.

4. The Test Drive: The h-BCN Material

To test their new tool, they used a material called h-BCN (a mix of Boron, Carbon, and Nitrogen arranged in a honeycomb pattern, like graphene).

• The Challenge: They wanted to find the arrangement of atoms that made the material the most stable (lowest energy).
• The Result: The "Old Language" (One-Hot) found a good solution, but it was slow and sometimes got stuck. The new "Modified Neighbor-Atom" (NAmod) language found the best solution much faster and more reliably. It was like switching from a map with just street names to a map that also shows traffic patterns and one-way streets.

Why Should You Care?

This might sound like abstract science, but it's the key to accelerating discovery.

• Better Batteries: Finding the perfect atomic arrangement could lead to batteries that charge in seconds and last for years.
• Superconductors: It could help design materials that conduct electricity with zero resistance at room temperature, revolutionizing power grids.
• Faster Innovation: Instead of waiting years for a supercomputer to try every possibility, tools like PyAPX can guide scientists to the "golden ticket" in a fraction of the time.

In a nutshell: PyAPX is a smart assistant that helps scientists stop guessing and start knowing exactly how to arrange atoms to build the materials of the future. It's not just about what the atoms are, but how they hold hands.

Technical Summary

1. Problem Statement

While materials discovery has advanced significantly in Crystal Structure Prediction (CSP) (finding stable structures for a given composition) and Elemental Substitution (finding optimal compositions for a given structure), a critical gap remains: Atomic Configuration Exploration.

Even when the crystal lattice and chemical composition are fixed, material properties (e.g., band gaps in 2D materials) can vary drastically depending on the specific arrangement of atoms at crystallographic sites. Existing tools often rely on random sampling (like Special Quasirandom Structures) or lack user-friendly interfaces to couple Bayesian optimization with first-principles calculations (DFT) for targeted configuration searches. There is a need for a toolkit that can efficiently navigate the combinatorial space of atomic arrangements to find the most stable or desirable configurations.

2. Methodology: PyAPX Workflow

The authors introduce PyAPX, a Python toolkit designed to automate the search for stable atomic configurations by integrating Bayesian Optimization (BO) with Density Functional Theory (DFT).

A. Workflow Architecture

The workflow consists of two main stages:

1. Pre-processing: The user defines a search space by providing a list of candidate atomic configurations. These are transformed into feature vectors (encodings).
2. Main Loop (Bayesian Optimization):
   • Sampling: A Bayesian model (implemented via the PHYSBO library) predicts the DFT total energy and uncertainty for candidate structures based on acquired data.
   • Acquisition: An acquisition function (e.g., Thompson sampling) balances exploration (sampling uncertain regions) and exploitation (sampling promising regions) to select the next structure.
   • Evaluation: The selected structure is passed to a DFT code (currently supporting Quantum ESPRESSO) to calculate the total energy. The result is fed back to the model to update the surrogate.
   • Iteration: This loop repeats until a specified number of iterations is reached.

B. Novel Encoding Methods

A core technical contribution is the development of encoding schemes specifically designed to capture local atomic environments, which standard methods often miss. The paper compares three encoding strategies:

1. One-hot Encoding: The standard approach where each site is represented by a vector indicating the presence of specific elements (1 for present, 0 otherwise). Limitation: It ignores spatial relationships between sites.
2. Neighbor-Atom (NA) Encoding: Modifies one-hot encoding by adding a weighted contribution from neighboring sites.
   • Formula: $\phi_{NA} = \text{one-hot} + w \times \text{neighbor\_counts}$.
   • This incorporates local convolution of atomic occupancy.
3. Modified Neighbor-Atom (NAmod) Encoding: Further enhances NA encoding by incorporating local anisotropy.
   • It calculates the variance ($\sigma^2$) of the NA vectors among neighboring sites to capture directional differences in the local environment.
   • Formula: $\phi_{NAmod} = [\phi_{NA}, \sigma]$.
   • This allows the model to distinguish between symmetric and asymmetric local environments, which is crucial for complex alloys.

3. Key Contributions

• PyAPX Toolkit: A user-friendly, open-source Python toolkit that bridges Bayesian optimization libraries with DFT codes, specifically tailored for atomic configuration pattern exploration.
• Advanced Encoding Schemes: The proposal and validation of NA and NAmod encoding methods. These methods embed richer topological information (local environment and anisotropy) into the feature vectors compared to traditional one-hot encoding.
• Demonstration on h-BCN: A comprehensive benchmark using a hexagonal Boron-Carbon-Nitride (h-BCN) system, a 2D material where properties are highly sensitive to atomic arrangement.

4. Results

The authors evaluated the performance of the four encoding methods (One-hot, NA, NAmod, and NAmod with PCA dimensionality reduction) on a $(3\times3)$ periodic h-BCN system (18 sites, 6 B, 6 C, 6 N atoms).

• Convergence Speed:
  • One-hot: Showed effective optimization but converged slower.
  • NA: Performed similarly to or slightly worse than one-hot in terms of cumulative minimum energy, suggesting simple neighbor convolution is insufficient for this system.
  • NAmod: Demonstrated superior convergence. It achieved lower cumulative minimum energies and faster reduction in moving averages compared to both one-hot and NA encodings.
  • NAmod + PCA: Dimensionality reduction (reducing 72D to 54D) maintained the performance of NAmod and showed even faster convergence in cumulative minimum, likely due to reduced noise and improved model training efficiency.
• Stability Identification: The NAmod approach successfully identified four symmetry-equivalent most stable configurations that were missed or found later by other methods.
• Robustness: Five independent experiments confirmed that NAmod and NAmod+PCA consistently outperformed one-hot encoding, validating that the improvement is due to the encoding quality rather than stochastic factors.

5. Significance

• Advancing Materials Design: PyAPX addresses a critical bottleneck in materials discovery: the "configuration problem." It enables researchers to move beyond finding a stable structure to finding the optimal atomic arrangement for specific properties.
• Efficiency: By using Bayesian optimization with superior encodings, the toolkit significantly reduces the number of expensive DFT calculations required to find global energy minima in combinatorial spaces.
• General Applicability: While demonstrated on h-BCN, the toolkit is designed for general crystalline materials. The NAmod encoding is particularly promising for systems where local anisotropy dictates stability, such as solid solutions and surface reconstructions.
• Accessibility: By providing a ready-to-use Python interface, PyAPX lowers the barrier for integrating machine learning-driven configuration searches into standard computational materials science workflows.

In conclusion, the paper establishes that incorporating local environmental anisotropy into feature encoding (NAmod) is a critical factor for efficient Bayesian optimization of atomic configurations, and PyAPX provides the practical infrastructure to leverage this insight for accelerated materials discovery.